AWS Knowledge Base

There is no generic @modelcontextprotocol/server-aws package. The official scope only ever shipped a Knowledge Base retrieval server, which now sits in servers-archived. If you need broad AWS operations — EC2, S3, CloudWatch — look at the community aws-mcp family instead; this entry is about KB retrieval specifically. What it does: - Exposes a single retrieve tool parameterised by Knowledge Base ID - Calls Bedrock Agent Runtime's Retrieve API and returns the top passages - Uses the host's AWS credential chain — env vars, shared profile, EC2/IRSA - Returns source URIs alongside each passage so the agent can cite grounding - Accepts a numberOfResults arg so the agent can trade recall for precision - Does no provisioning, no IAM, no compute — it is a read-only search tool Example: Point it at a KB that ingests your internal runbooks and wiki. An SRE agent answering "how do we roll back the billing service" retrieves three ranked passages with their S3 source URIs, then composes the answer with real citations instead of hallucinating a procedure. Setup: AWS_REGION plus credentials via any mechanism the AWS SDK for JavaScript recognises. KNOWLEDGE_BASE_ID identifies the KB to query. The principal needs bedrock:Retrieve on the target KB ARN. Transport: stdio. Notes: Because this is an archived reference, it ships only KB retrieval — not KB management, not data-source sync. Those are separate Bedrock APIs and require their own tooling. Reliability score is lower here to reflect archival status, not code quality. For production KB workflows, pair this with a Fetch or filesystem server so the agent can also read the cited sources directly when a passage is insufficient.